Re: Interesting scheduling times - NOT

Richard Gooch (rgooch@atnf.csiro.au)
Fri, 25 Sep 1998 10:24:11 +1000

Larry McVoy writes:
> Richard Gooch <rgooch@atnf.csiro.au>:
> : > The issue is that Richard's variance is so high.
> :
> : On another set of lmbench runs (no low-priority processes), I got:
> : "size=0k ovr=6.23
> : 2 5.12
> : 2 3.97
> : 2 3.96
> : 2 3.94
> :
> : yet a second later I got:
> : "size=0k ovr=6.23
> : 2 5.12
> : 2 5.12
> : 2 5.13
> : 2 4.35
> :
> : so it's not hard to find a 30% variance in lmbench either.
>
> That's weak. You'll be hard pressed to find a consistent variance
> in that benchmark. Go run it 10 times in a row and tell us what the
> variance is. I just did it and got 12%. On the other hand, you've been
> telling us all along that your tests varies 70-100% every run. If you
> had started out the conversation: I ran ten times and go one case that
> was off by a 100%, but the rest all nicely clustered around one point,
> I don't think anyone would have had an issue, now would they?

I don't recall saying anywhere that my times vary on every run. I'm
hard pressed to find a consistent pattern to the variance. I can go
for dozens of runs and get numbers within a few percent, and then
suddenly I get one or two runs with a higher
minimum/median/average/maximum after which it goes back down to the
lower numbers. Other times I get long periods where I usually get the
higher numbers, and then suddenly the numbers drop down again.

I've looked for correlations with the time of day (possible cron jobs
or other daemon activity). I've even resorted to killing off processes
one by one to no avail, until I'm left with init, a few gettys and the
update/bdflush daemons. No difference.

I can leave my machine for an hour and come back and get the same
results as the last run. Later, after coding, compiling and generally
working the machine, I may start to see different numbers.

About the same time I was measuring variance from lmbench, I was
getting reasonable stability from my test.

OK, you wanted 10 runs. I've appended them. They were done in quick
succession, spaced about 0.5 s apart. Just before and after running
lat_ctx I ran my benchmark (with 10 extra processes) and got stable,
low variance results beforehand. Afterwards, I got a couple of runs
with normal values, a few runs with significantly larger values, and
then it went back down to where it normally sits.

Another idea I've got at the moment is the FPU state saving (yes, I do
have the fix Linus sent out for 2.1.122), so I'm exploring that.

Regards,

Richard....

% lat_ctx 2 2 2 2

"size=0k ovr=6.23
2 5.18
2 5.17
2 5.16
2 4.08
% lat_ctx 2 2 2 2

"size=0k ovr=6.23
2 5.25
2 5.25
2 5.25
2 5.24
% lat_ctx 2 2 2 2

"size=0k ovr=6.21
2 5.16
2 5.14
2 3.98
2 3.98
% lat_ctx 2 2 2 2

"size=0k ovr=6.23
2 5.29
2 5.29
2 5.30
2 5.30
% lat_ctx 2 2 2 2

"size=0k ovr=6.23
2 5.18
2 5.18
2 5.18
2 5.18
% lat_ctx 2 2 2 2

"size=0k ovr=6.23
2 5.25
2 5.25
2 5.25
2 5.25
% lat_ctx 2 2 2 2

"size=0k ovr=6.23
2 5.13
2 3.96
2 3.95
2 3.94
% lat_ctx 2 2 2 2

"size=0k ovr=6.23
2 5.29
2 5.29
2 4.03
2 4.01
% lat_ctx 2 2 2 2

"size=0k ovr=6.23
2 5.17
2 5.18
2 4.64
2 4.08
% lat_ctx 2 2 2 2

"size=0k ovr=6.23
2 5.26
2 5.22
2 5.21
2 4.03
%

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/